memnest

English | 한국어

Your coding AI forgets everything when the session ends. Memnest lets it look up what you decided before, and why, from the original conversation instead of an AI-written summary.

• Original text, not summaries. Saving and searching call no model API, so there is no API key and no extra token bill. Every result has a source ID you can open.

• One memory for every coding AI. Claude Code, Codex, Cursor, and pi share one local service. Decide something in Claude Code, find it from Codex.

• Stays on your machine. One Rust binary with SQLite and a local multilingual embedding model.

Install

curl -fsSL https://raw.githubusercontent.com/Blue-B/memnest/main/core/scripts/install.sh | bash

On Windows, in PowerShell (no administrator rights needed):

irm https://raw.githubusercontent.com/Blue-B/memnest/main/core/scripts/install.ps1 | iex

This starts the server and the conversation watcher, and connects the Claude Code, Codex, and Cursor setups it finds. Existing configuration is backed up first, and the restore command is printed. Codex runs the new hooks after you trust them once with /hooks.

Needs macOS on Apple Silicon, Linux with systemd (x86_64 with glibc 2.35+, Arm64 with glibc 2.39+), or 64-bit x86 Windows 10/11, and Python 3. On Windows, Python is only used to connect clients (details). The first write or search downloads an embedding model of about 1.1 GB.

• pi: pi install npm:pi-memnest@0.6.0, then check with /memnest.

• Other MCP clients: connect to http://127.0.0.1:3111/mcp (examples).

Related MCP server: agent-history-mcp

Use it

Ask in plain words. Your AI picks the tool.

A search result is the stored text with its project, score and ID:

[1] project=shop-api score=0.2531 id=manual_5ea65b71... doc_len=96
    Deploy from the release branch only, never on Fridays. Decided 2026-10-02 after the Friday rollback.
hint: stronger matches exist outside this workspace scope: billing (2). Search again with project="billing" (or project="all").

The hint line appears when another project holds clearly better matches than the folder you are in. It names the project and a count, never the text.

Tools: memory_search, memory_get, memory_remember, memory_update, memory_delete, task_card.

How it works

 Claude Code      Codex      Cursor      pi        ChatGPT, Claude, Grok (web)
   MCP + hook   MCP + hook    MCP     extension    read-only gateway (optional)
       |            |          |          |                    |
       +------------+----------+----------+--------------------+
                                |
                 memnest server (one Rust binary, 127.0.0.1:3111)
                   save       text stored as written, embedded locally
                   search     keyword (BM25) + vector, fused and ranked
                   context    memories on the prompt's topic, for the hook
                   task cards, vault
                                |
                 ~/.memnest
                   memory.db    SQLite, the source of truth
                   text_index   keyword index (Tantivy)
                   vectors      vector index
                   models       multilingual-e5-base, runs on the CPU
                   master.key   key for the vault (AES-256-GCM)

 memnest watch: reads the session files of Claude Code, Codex and pi
                and stores each visible turn through the same server

1. Save. A memory is stored exactly as it was written. Nothing rewrites or summarizes it. A small local model turns the text into numbers (an embedding) so it can be found by meaning later.

2. Search. Two searches run on the same question: keyword search finds the same words (file names, numbers, exact terms), vector search finds the same meaning in different words. The two rankings are merged, memories you saved on purpose rank above raw conversation turns, older ones rank lower, and near-duplicates are thinned out. Three results come back by default, each with an ID.

3. Automatic recall. With the prompt hook, memnest looks up memories when you send a prompt and adds only the ones on that prompt's topic. A vague prompt gets none.

4. Scope. Each folder you work in is its own workspace. playbook is shared by all of them. Credentials go to the encrypted vault and never appear in search.

No language model runs in any of these steps, which is why saving and searching need no API key. Design reasons: design decisions. How search ranks and what it was measured on: benchmarks.

Task cards

For work that runs over several days. Your AI writes the card, Memnest adds your own messages to it word for word, and the next session opens the card and continues. In a replay of a real five-day pi session, the reply read 174K context tokens instead of 321K. Your stated policy was respected in 3 of 3 blind-graded runs with the card, against 0 of 3 without it. A card does not force the model to follow it. How it works and limits

Commands in pi

Type these in pi to look at memnest yourself. They run without the model and add nothing to its context.

/memnest-review is there because memnest cannot tell whether an old rule still holds; you can.

Also included

Conversation capture (memnest watch), history import, code evidence, an encrypted vault for credentials, team access, and web connectors for ChatGPT, Claude, and Grok. Automatic recall is off unless you turn it on.

Limits

• Search can miss a record, and it cannot tell whether an old decision still holds. When your AI saves a memory, the reply lists existing memories that may cover the same subject, so it can replace the old one instead of adding a second copy. On two public test sets, measured without any language model, the conversation that holds the answer is among the first 5 results for 98% of questions (LongMemEval-S, all 470 questions that have an answer; released 0.5.1: 95%), and the exact line that holds the answer is among the first 10 for 80% (LoCoMo; 0.5.1: 70%). These count whether the memory was found, not whether the AI then answered correctly. What was tested and how to read it

• Capture keeps visible conversation text only: no reasoning, tool output, or images.

• Memories your AI reads are sent to its provider. Regular memories are not encrypted, so do not expose port 3111 to the internet; a web connector URL works like a password.

• Deleted records stay in trash for 30 days, and backups may keep copies. See security.

More

Operations, upgrades, and backups, changelog, contributing

License

MIT © Blue-B